Newsgroups: comp.parallel.pvm
Path: ukc!uknet!doc.ic.ac.uk!agate!howland.reston.ans.net!cs.utexas.edu!asuvax!newncar!csn!hellgate.utah.edu!lanl!newshost.lanl.gov!dwf
From: dwf@acl.lanl.gov (David W. Forslund)
Subject: Re: PVM on CM5....
In-Reply-To: alan@af.msc.edu's message of 12 Jan 1994 03:11:59 GMT
Message-ID: <DWF.94Jan14165502@coyote.acl.lanl.gov>
Sender: news@newshost.lanl.gov
Reply-To: David W. Forslund <dwf@acl.lanl.gov>
Organization: Advanced Computing Lab, LANL, NM
References: <peschko.758261197@s13> <2gvppv$4m0@uc.msc.edu>
Date: Fri, 14 Jan 1994 23:55:02 GMT
Lines: 60

In article <2gvppv$4m0@uc.msc.edu> alan@af.msc.edu (Alan Klietz) writes:

   Path: lanl!hellgate.utah.edu!dog.ee.lbl.gov!agate!howland.reston.ans.net!math.ohio-state.edu!darwin.sura.net!news-feed-1.peachnet.edu!umn.edu!msc.edu!af.msc.edu!alan
   From: alan@af.msc.edu (Alan Klietz)
   Newsgroups: comp.parallel.pvm
   Date: 12 Jan 1994 03:11:59 GMT
   Organization: Minnesota Supercomputer Center, Inc.
   Lines: 33
   References: <peschko.758261197@s13>
   NNTP-Posting-Host: af.msc.edu

   In article <peschko.758261197@s13> peschko@arc.umn.edu (Ed Peschko) writes:
   <hello all --
   <
   <How does PVM stand up for implementation on the CM5 compared to the distributed
   <computing libraries *inherent* on the CM5  itself? (I am thinking CMMD here).
   <Has anyone experimented with it / compared the two? And does anyone have any 
   <general pointers on PVM's use with the CM5?
   <
   <Thanks much,
   <
   <Ed

   My understanding is that the CM-5 implementation of PVM 3.x sits on top
   of CMMD, i.e., it thunks down to CMMD, which makes the actual communication 
   calls to the Network Interface chip.

   Of course, whenever you do translation to an underlying API, you take
   a performance hit relative to just calling the API directly.  I don't know
   of any formal studies of PVM-vs-CMMD in particular.

   As a pure guess, because PVM requires an extra round trip to memory for
   the canonicalization buffer, I'd speculate a 20%-30% bwidth hit.  Future 
   versions of PVM will implement PvmDataInPlace, passing a pointer to CMMD 
   instead of copying the data.  Then the only overhead should be in the 
   repackaging of the argument parameters.

   --
   Alan E. Klietz
   Minnesota Supercomputer Center, Inc.
   1200 Washington Avenue South
   Minneapolis, MN  55415
   Tel: +1 612 337 3520	       Internet: alan@msc.edu
   Fax: +1 612 337 3400	

Some of our folks here who have used it indicate that the hit is more
like 3-4 times in some cases.  I don't know the reasons for this,
however.

 
--
David Forslund
Advanced Computing Laboratory
MS B287
Los Alamos National Laboratory
Los Alamos, NM 87545

Voice:(505) 665-1907
FAX: (505) 665-4939
EMAIL: dwf@lanl.gov

